FranckQC commented on pull request #9482:
URL: https://github.com/apache/tvm/pull/9482#issuecomment-1031887221


   Well, to be honest, I'd love to see this pass as a default as it's a 
target-independent optimization from which everybody could benefit without 
having to wonder weither the back-end is going to take care of redundant 
computations. I see optional passes as more specialized or more target 
dependent.
   
   I also believe that our TVM CSE it is more aggressive than the LLVM CSE, 
which can't deal with things like divisions because they face some restrictions 
for semantics preservation that are not as strong in our specific DSL case.
   
   As that was only 4 or 5 tests that were annoying, I think we should still be 
able to make it a default pass. Most of them I could already fix with just 
disabling the CSE pass (mlost of them were in test_lower_build.py, which was 
expected). Another possibility would have been to rewrite the test and to state 
the equality between the new expected and what we got. But I think it's better 
to be minimalist, and to avoid running other passes that "complicate" the 
output.
   
   I'll try to fix the last two issues and we should be good.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to